HW2.5 capabilities

[BigD0g]

Tesla's focus seems to be on the Tesla Network. They are trying to make this work so they can start making some money. I think it will be something very cool because when you get into the car, the car will not just take you where you want to go driverlessly but it will also be lots of fun because of the personalization settings. On the continental website, the following was the important feature about the internal camera:



What they are describing here matches this recent Electrek article:
Tesla is building more features geared for car-sharing, now cloud-based driver profiles are coming, says Musk

Yes, and by the time the S/X have it, we'll be talking about HW3.0

 

[scottf200]

We don't know yet which Conti model, but a quick compare between the Bosch MRRevo14 (existing Tesla radar) and the "Long Range" Conti models indicates that the new radar will have much longer detection range: 220-250 m. vs Bosch's 160 m.

Some spec differences

Bosch

ARS441

ARS510

 

[number12]

If 2.5 was enough for FSD they would have just called in 3.0. They either don't want to caniblize sales in the short term, or are not confident in HW yet.

 

[BigD0g]

If 2.5 was enough for FSD they would have just called in 3.0. They either don't want to caniblize sales in the short term, or are not confident in HW yet.

I'm going to go with no confident in the HW, they clearly don't have the software working, so they can't possible know what hardware is required to run it.

 

[MXZoomZoom]

I spoke to someone a year ago on the Autopilot team. They were not super confident that 2.0 hardware was going to get it done (L5 proper). And it looks like that's indeed the case.

 

[Bebop]

Yes, and by the time the S/X have it, we'll be talking about HW3.0

The promise is that 2.0 is FSD capable. They keep selling that promise even as an option when you buy a new Tesla right now. You are right that obviously like an iPhone, tech keeps improving, so yes I'm sure there will be a 3.0 at some point.

 

[pilotSteve]

The promise is that 2.0 is FSD capable. They keep selling that promise even as an option when you buy a new Tesla right now. You are right that obviously like an iPhone, tech keeps improving, so yes I'm sure there will be a 3.0 at some point.

especially for those like me that already purchased FSD. Oddly enough I had a premonition that the hardware would likely require several upgrades (not just firmware) and I'm going to hold Tesla to delivering what they promised without paying more than I already have.

 

[Bebop]

especially for those like me that already purchased FSD. Oddly enough I had a premonition that the hardware would likely require several upgrades (not just firmware) and I'm going to hold Tesla to delivering what they promised without paying more than I already have.

I'm in the same exact position, although I haven't taken delivery of my car yet. I'm not worried at all. If upgrades need to happen, I'm more than certain Tesla will take care of the situation. Otherwise here come the lawsuits...

 

[Alketi]

Why are we assuming the internal camera is based on Continental's tech and not nVidia's Co-Pilot when nVidia is already providing the computer that it would connect to?

 

[Bladerskb]

It's not going to compensate. Tesla never said anything like that. They only confirmed that AP2.0 will be twice as safe as the average human driver. Check out the following quote:

Some people read this statement as "if it turns out AP2.0 is not as good as AP2.5...". However, that's not what they said. The criteria for AP2.0 is 2x safety. Tesla is trying to achieve 10x safety. Elon talked about 10x many times. AP2.5 won't achieve that either.

The only people who regurgitate the 2x safety talking point is the ppl who buy the tesla pr campaign.

You do realize that elon has already gone around saying that autopilot (ap1) was already 2x better than human?

Making the whole 2x safety irrelevant since they can throw out same statistic as ap1 and sat ap2 fsd is 2x.

What really matters is disengagement per mile.

My best guess is, AP2.0 will have FSD with the driver present in the city and on the highway. AP2.5 will have driverless FSD in the city but driver present on the highway. By FSD I mean entering an address and the car taking you there.

That is quite incorrect, city driving is 1000000x magnitude more difficult than highway driving.

 

[Troy]

That is quite incorrect, city driving is 1000000x magnitude more difficult than highway driving.

I have done some calculations and I don't think the EAP (AP2.0 or AP2.5) hardware can do driverless lane changes on the highway safely because it can't detect cars approaching from behind far enough at high speeds. In fact, I don't think Tesla is even trying to do that because unlike Mercedes, Tesla doesn't have a rear radar and even Mercedes doesn't try to do it even though the rear radar has more range than Tesla's cameras.

However, I think the Model 3 will do driverless lane changes in specific situations. To understand this topic better, the question you have to ask yourself is this: "How can EAP be sure that a bullet car is not approaching from behind at high speed on the target lane the EAP wants to change to?"

I thought about this and I think the answer is, if there is already a car behind you, let's say 20 meters behind you on the target lane, then you can be sure there is no bullet car. Even if there was, it would hit this other car that is 20 meters behind you. I wrote about it in detail here.

In other words, EAP will do driverless lane changes not when the target lane is clear, but ironically, when the target lane is occupied but the car is far enough. Elon described this as "passing and maneuvering around other cars without touching anything with Enhanced Autopilot". Listen here. Therefore this will work better in the city.

Tesla's own diagram here shows the rear cameras can only see 100 meters and 50 meters behind. This diagram is still available on the Tesla website here below the video. In comparison, the Mercedes hardware shown here and here on their website can see 200 meters behind. However, my calculation shows that around 300-350 meters rear range is needed. Details here. Therefore Tesla is not even attempting to do driverless lane changes on the highway.

 

[Alketi]

In other words, EAP will do driverless lane changes not when the target lane is clear, but ironically, when the target lane is occupied but the car is far enough. Elon described this as "passing and maneuvering around other cars without touching anything with Enhanced Autopilot". Listen here. Therefore this will work better in the city.

Troy, I don't think your conclusions necessarily follow from your math nor do they prove that highway lane-changing will be extraordinarily difficult.

1. You're assuming max closing speed.
2. You're assuming theoretically perfect safety.
3. You correctly state that lane changes are safer when the target lane is occupied.
4. But then conclude that it will work better in the city.

AP does not need perfect (100%) safety to be drastically better than humans and save lives, nor will it require that guarantee in order for it to be road-worthy.

An AP car being rear-ended by a 155MPH human-guided missile will not lead to calls for AP to be banned but rather for human drivers to be removed as the real safety hazards ...

An AP car could easily wait for adjacent lane occupancy within it's max rear vision radius for a specified time or for occupancy 25m ahead before changing lanes. This means there would be a "blocking car" in either the rear or forward positions. There's nothing difficult about that set of limitations.

City lane changes require all sorts of random merging and u-turn avoidance not present on a highway.

 

[lunitiks]

How do you guys define "better than humans", "2x safety" etc. I mean in what context? Are you talking about, like, overall crash stats disregarding road type, population and other variables? Would it be OK (politically) to accept a dangerous auto lane change system in certain scenarios as long as it excels on others, given that it statistically reduces the overall accident rate? What's "perfect", "better" and "good enough"? Am I complicating this too much?

I don't mean to derail this thread so maybe there's some other appropriate thread for this particular discussion...

 

[Pale_Rider]

Never mind that if a Tesla initiates a lane change and something comes into view that it can't account for it can always abort the lane change and put you back where you were. There are lots of ways to compensate for a car going too fast relative to your velocity when changing lanes. We do it all the time when we drive. Aborting a lane change and speeding up are two... Some how they manage on the autobahn, after all, not everyone is driving 155...

 

[Troy]

How do you guys define "better than humans", "2x safety" etc. I mean in what context?

I think they will use airbag deployments as the criteria.

Elon: The probability of having an accident is 50% lower if you have Autopilot on. Even with our first version. So we can see basically what’s the average number of kilometers to an accident – accident defined by airbag deployment. Even with this early version, it’s almost twice as good as a person. Source

 

[Pale_Rider]

I can run over lots of pedestrians and bikes and never make an airbag deploy...lol

 

[lunitiks]

Airbag deployments pr mile seems like a pretty bad way to measure auto lane-change safety/capability, though...

 

[aWalkingShadow]

I think they will use airbag deployments as the criteria.

But the question is whether they're comparing the same situations. Autopilot is mainly used on highways. Is the control group, cited by Tesla, cars driving on highways, or is it all cars?

 

[FlatSix911]

WSJ casts negative light on Tesla’s Autopilot ambitions

The publication claims that CEO Elon Musk’s ambitious goals for the program have led to some employees resignations and doubts about the company’s ability to deliver on those goals. Based on unnamed sources, WSJ claims that the decision to say that Autopilot 2.0 will enable full self-driving capability wasn’t well received: “In a meeting after the announcement, someone asked Autopilot director Sterling Anderson how Tesla could brand the product “Full Self-Driving,” several employees recall. “This was Elon’s decision,” they said he responded. Two months later, Mr. Anderson resigned.”
The rest of the report revolved more around the first generation of the Autopilot and hurdles to bring the technology to market. They claim that the Autopilot team lost “at least 10 engineers and four top managers” in recent months. We reported on several computer vision experts leaving the group in June and software executive Chris Lattner left a month before.

 

[AnxietyRanger]

HW 2.5 is not just a CPU/GPU change (an easy upgrade for HW 2.0), nor is it a generational shift (e.g. HW 3.0 with twice the sensors or something), which suggests to me HW 2.5 is actually plugging holes in HW 2.0 with redundant wiring, new radar, new chips and power elements here and there.

I.e. HW 2.5 is probably the worst kind of vote of no-confidence in HW 2.0 given how soon it appeared. HW 2.5 seems like Tesla covering their behind. But will it be enough?

HW 2.5 seems like P90DL V3 to HW 2.0's P90DL V1.